Graph and medical record dual-source adaptive retrieval generation method and system for dizziness clinical decision

Abstract

The application provides a graph and medical record dual-source adaptive retrieval generation method and system for dizziness clinical decision, belonging to the technical field of medical information intelligent processing, and taking standardized clinical entities as anchor nodes to perform multi-hop path reasoning, output candidate reasoning paths with topology scores; parallel retrieval of previous medical records similar to patient complaint semantics, output candidate medical records with semantic similarity scores; standardization and entropy-driven dynamic gating fusion of the two scores to obtain an optimal evidence subset; structured assembly of the optimal evidence subset, patient metadata and system prompt templates to drive a large language model to output differential diagnosis suggestions with explicit evidence traceability. The application realizes parallel retrieval of structured knowledge and proven evidence and strictly scaled unified fusion in mathematics; realizes information-driven adaptive adjustment of fusion weights; and provides transparent evidence chains traceable to specific knowledge sources and medical records.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical information processing technology, specifically to a dual-source adaptive retrieval method and system for generating diagnostic opinions on dizziness and vertigo using both knowledge graphs and medical records. It particularly addresses the auxiliary differential diagnosis of dizziness and vertigo, integrating structured medical knowledge graph topological reasoning with unstructured electronic medical record semantic retrieval to generate dual-source adaptive diagnostic opinions. The method and system of this invention are applicable to auxiliary decision-making in vertigo clinics in clinical departments such as otolaryngology and neurology, and can also be deployed in the initial screening stage of vestibular disorders in primary healthcare institutions. This helps clinicians integrate structured medical knowledge with real-world clinical experience in the complex differential diagnosis of dizziness and vertigo, improving the accuracy and interpretability of the diagnosis. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, clinical decision support systems have played an increasingly important role in assisting doctors in diagnosis and treatment and reducing the misdiagnosis rate.

[0003] Dizziness and vertigo are among the most common clinical symptoms, accounting for approximately 4% to 5% of emergency room visits worldwide, with a prevalence exceeding 30% in the elderly population over 65 years of age. Differential diagnosis of dizziness and vertigo is a recognized challenge in clinical medicine. Benign peripheral causes, Meniere's disease, vestibular neuritis, vestibular migraine, and potentially life-threatening central causes highly overlap in clinical presentation. Furthermore, the diagnostic process heavily relies on the patient's verbal, non-standardized description of subjective symptoms (e.g., describing symptoms as "vertigo," "walking on cotton," or "lightheadedness") and the clinician's mastery of complex exclusionary logic. Therefore, the differential diagnosis of dizziness and vertigo is highly complex.

[0004] From a technological evolution perspective, the application of artificial intelligence in medical diagnosis has progressed from rule-based systems to machine learning, and then to deep learning and large language models. Early medical expert systems relied on manually coded rule bases, which were costly to maintain and had weak generalization capabilities. In recent years, the rapid development of large language models has opened up new paths for medical artificial intelligence—large language models, after pre-training on massive amounts of medical literature and clinical texts, have demonstrated powerful medical knowledge reserves and semantic understanding capabilities.

[0005] However, LLM faces a core bottleneck in clinical decision-making scenarios: the medical knowledge in the pre-training corpus is outdated and may contain errors, causing the model to produce seemingly reasonable but actually erroneous "illusion" outputs during reasoning. This is unacceptable for clinical decision-making scenarios with extremely high safety requirements. Specifically for the vertical field of dizziness and vertigo, existing technologies have the following significant shortcomings:

[0006] First, a significant semantic gap exists between colloquial complaints and standardized terminology. Patients with dizziness and vertigo often use colloquial descriptions of their symptoms—such as "feeling like the world is spinning," " feeling lightheaded," "walking on cotton," or "feeling unsteady on their feet"—but existing general-purpose word segmenters and vectorization methods cannot accurately capture the standardized clinical semantics behind these colloquial expressions. This causes subsequent retrieval and inference to deviate from the actual clinical state. Existing technologies, whether based on structured feature extraction from CED models or entity linking based on general graphs, lack a specialized semantic mapping mechanism for the vestibular medicine field.

[0007] Secondly, existing retrieval methods suffer from "semantic flattening" when dealing with complex medical scenarios, only capable of simple text similarity matching, and failing to utilize the structured correlation logic between symptoms, negative signs, time patterns, precipitating factors, and examination results involved in disease differential diagnosis. There is a lack of in-depth understanding of the topological correlation logic in disease differential diagnosis.

[0008] Furthermore, while some emerging graph retrieval enhancement technologies incorporate structured knowledge graphs as external knowledge bases, they often treat them as a single source of evidence. In actual clinical decision-making, experts rely not only on rigorous medical guidelines (structured knowledge) but also heavily on the rich clinical experience contained in historical real-world electronic medical records (EMRs). Existing technologies have failed to achieve synergy between dual-source evidence from knowledge graphs and unstructured real-world medical records. This makes it difficult for large models to dynamically balance "textbook logic" and "real-world experience" when generating diagnostic recommendations, making them highly susceptible to factual errors when faced with rare or atypical cases of vertigo. Summary of the Invention

[0009] The purpose of this invention is to provide a dual-source adaptive retrieval method and system for generating knowledge graphs and medical records for clinical decision-making in dizziness and vertigo. This method accurately understands patients' spoken complaints by mapping them to standardized clinical entities, and enhances the accuracy and reliability of auxiliary diagnosis of vertigo by synergistically utilizing knowledge graphs and real medical records. This addresses at least one of the technical problems existing in the background art.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a dual-source adaptive retrieval and generation method for atlas and medical records in clinical decision-making for dizziness and vertigo, comprising:

[0012] Mapping patients' verbal complaints to a standardized set of clinical entities;

[0013] Based on the standardized clinical entity set, multi-hop path reasoning is performed in the dizziness and vertigo-specific knowledge graph with each entity as an anchor node to obtain candidate reasoning paths and calculate the corresponding topology scores.

[0014] Parallel retrieval of past medical records semantically related to the patient's chief complaint and the standardized clinical entity set is performed to obtain candidate medical records and calculate the corresponding semantic similarity score;

[0015] The topological score and the semantic similarity score are standardized respectively; dynamic fusion weights are determined based on the information entropy of the graph-side evidence set and the information entropy of the medical record-side evidence set, and a comprehensive fusion score is calculated accordingly; the optimal evidence subset is selected based on the comprehensive fusion score.

[0016] The optimal subset of evidence is structurally assembled with patient metadata and system prompt templates, driving the large language model to perform chain reasoning and output differential diagnosis suggestions with explicit evidence tracing.

[0017] As a further limitation of the first aspect of the present invention, a dizziness-specific knowledge graph is pre-constructed, wherein the dizziness-specific knowledge graph is a directed heterogeneous attribute graph, and its node entity types include at least the following five categories: positive symptom nodes, negative sign nodes, attack time pattern nodes, inducing factor nodes, and physical examination discovery nodes; the directed edge relationship types include at least the following five categories: diagnostic relationship, exclusion relationship, suggestive relationship, accompanying relationship, and possibly related relationship; wherein, the clinical weight value of the exclusion relationship is greater than the clinical weight value of the accompanying relationship, so as to reflect the priority of exclusion logic in differential diagnosis.

[0018] As a further limitation of the first aspect of the present invention, the patient's spoken complaint text is obtained, and the spoken complaint text is mapped and standardized into a standardized set of clinical entities through an entity linking module based on a pre-set vertigo specialist dictionary, including:

[0019] Receive the patient's natural language complaint text, which is the patient's original statement of dizziness and vertigo-related symptoms in a colloquial manner; slide the complaint text. Window scanning obtains a set of text fragments, among which The window length is set from 1 to a preset maximum length to accommodate the variability in phrase length in Chinese medical terminology; each text segment is vectorized using a dense embedding model in the medical field to obtain segment embedding vectors; a pre-built specialty clinical dictionary is obtained, which includes... A set of standard vestibular medical terms is defined, and each dictionary term is vectorized using the dense embedding model of the medical field. For each text segment and each dictionary term, a semantic alignment score is calculated. All terms that meet the criteria are extracted to form a standardized set of clinical entities.

[0020] As a further limitation of the first aspect of the present invention, based on the standardized clinical entity set, a parallel dual-source knowledge retrieval mechanism is triggered, including: retrieving topological evidence paths related to the standardized clinical entity set in a pre-constructed dizziness-specific knowledge graph; in the pre-constructed dizziness-specific knowledge graph, performing multi-hop traversal along directed edges with each entity as the starting anchor node to obtain a set of candidate reasoning paths, and calculating a topological score for each reasoning path; retrieving clinical medical record reference texts semantically similar to the standardized clinical entity set and the chief complaint text in an unstructured electronic medical record database; and using the weighted concatenation result of the mean pooling result of the embedding vectors of each entity and the embedding vector of the original chief complaint as the query vector in a pre-constructed electronic medical record vector database, obtaining the semantically most similar medical record through approximate nearest neighbor retrieval, and calculating a semantic similarity score.

[0021] As a further limitation of the first aspect of the present invention, the context of the fused prompt words is input into a pre-trained large language model. Through generative reasoning of the large language model, auxiliary diagnostic results and exclusive differential evidence interpretations for dizziness and vertigo are output, including: textualizing the graph reasoning paths and similar medical records in the optimal evidence subset; structurally assembling the diagnostic system prompt template, standardized patient metadata, graph reasoning path text, and similar medical record summaries and inputting them into the large language model to drive the model to perform reasoning, analyze the evidence segment by segment, compare with clinical manifestations, and perform differential exclusion; and output differential diagnostic suggestions with explicit evidence tracing and their reasoning tracing.

[0022] As a further limitation of the first aspect of the present invention, an adaptive gating mechanism is used to dynamically calculate the fusion ratio weight between the atlas retrieval results and the medical record retrieval results based on the atlas confidence score and the medical record confidence score, and to construct a fusion prompt word context based on the fusion ratio weight. This includes: calculating the Shannon information entropy of the retrieved atlas subgraphs, calculating dynamic gating parameters based on the information entropy and the entropy of the medical record retrieval results, calculating a comprehensive fusion score, sorting the results from high to low, and selecting... These pieces of evidence constitute the optimal subset of evidence.

[0023] Secondly, this invention provides a dual-source adaptive retrieval and generation system for atlases and medical records for clinical decision-making in dizziness and vertigo, comprising:

[0024] The dictionary-driven medical entity linker module is used to map patients' spoken complaints to a standardized set of clinical entities;

[0025] The Dizziness and Vertigo Specialized Knowledge Graph and Multi-hop Path Reasoning Engine Module is used to perform multi-hop path reasoning with standardized clinical entities as anchor nodes and output candidate reasoning paths with topological scores.

[0026] The electronic medical record vector database and dense vector retrieval engine module are used to retrieve past medical records that are semantically similar to the patient's chief complaint in parallel, and output candidate medical records with semantic similarity scores.

[0027] The adaptive Z-score gated fusion layer module is used to standardize the atlas-side score and the medical record-side score respectively, and perform dynamic gated fusion based on the information entropy of the evidence sets on both sides to output the optimal subset of evidence.

[0028] The evidence tracing generation layer module is used to structurally assemble the optimal subset of evidence with patient metadata and system prompt templates, drive the large language model to perform chain reasoning, and output differential diagnosis suggestions with explicit evidence tracing.

[0029] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the dual-source adaptive retrieval and generation method for clinical decision-making in dizziness and vertigo as described in the first aspect.

[0030] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo as described in the first aspect.

[0031] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the dual-source adaptive retrieval and generation method of atlas and medical records for clinical decision-making in dizziness and vertigo as described in the first aspect.

[0032] Terminology Explanation: RAG (Retrieval Enhanced Generation): Before generating an answer using LLM (Linguistic Modeling), evidence relevant to the current query is retrieved from an external trusted knowledge base, and this retrieved evidence is injected as context into the generation process. GraphRAG (Graph-based Retrieval Enhanced Generation): Introduces a knowledge graph into the RAG framework, utilizing structured graph path reasoning to enhance generation quality. KG (Knowledge Graph): A network of medical entities and their semantic relationships organized in a directed graph structure. EMR (Electronic Medical Record): De-identified electronic records of real-world patient medical records. LLM (Large Language Model): A generative language model based on large-scale text pre-training. Z-score: A statistical method for mapping data to a standard normal distribution with a mean of 0 and a standard deviation of 1. BPPV (Benign Paroxysmal Positional Vertigo): The most common cause of peripheral vertigo, caused by otoliths dislodging and entering the semicircular canals. MD (Meniere's Disease): An inner ear disease characterized by episodic vertigo, fluctuating hearing loss, tinnitus, and a feeling of fullness in the ear. VN (Vestibular Neuropathy): A viral infection of the vestibular nerve characterized by acute, persistent vertigo. VM (Virtual Vestibular Migraine): A central vestibular disorder characterized by paroxysmal vertigo accompanied by migraine. ICD-11 (International Classification of Diseases, 11th Revision): The internationally accepted classification of diseases published by the World Health Organization. Chain-of-Thought: A technical paradigm for analyzing reasoning step-by-step using large language models.

[0033] The beneficial effects of this invention are as follows: It eliminates the semantic gap between colloquial complaints and standardized terminology: Through a medical entity linking mechanism driven by a specialized clinical dictionary, colloquial expressions are accurately mapped to standard terminology, significantly improving the accuracy of understanding real-world patients' natural language input compared to existing technologies that rely on general word segmenters or predefined feature tables. It achieves parallel retrieval and mathematically rigorous scale unification of structured knowledge and empirical evidence: By standardizing Z-scores, heterogeneous scores are uniformly mapped to a standard normal distribution, eliminating the distribution scale difference between topological scores and vector similarity scores, allowing both types of evidence to contribute fairly to the final diagnostic decision. Compared to linear weighting or simple concatenation in existing technologies, it fundamentally avoids the systematic bias of "empirical noise drowning out knowledge facts." It achieves information-driven adaptive adjustment of fusion weights: The information density of the graph subgraph is calculated using Shannon information entropy, and the gating parameters are dynamically adjusted accordingly. —When the graph reasoning path is dense and information-rich, its weight is automatically increased; when the graph path is sparse, it automatically relies more on medical record evidence. Compared to the fixed weights used in existing technologies, this adaptive mechanism allows the system to maintain the optimal evidence fusion ratio under various query conditions. A dizziness-specific knowledge graph with specialized exclusionary diagnostic logic has been constructed: Unlike general medical knowledge graphs, the specialized graph of this invention is specifically optimized for the differential diagnosis needs of vestibular medicine in terms of entity granularity and relationship types, especially the structured modeling of key elements of exclusionary diagnosis such as negative signs and time patterns, enabling the graph reasoning path to truly capture complex clinical exclusionary logic. A transparent evidence chain traceable to specific knowledge sources and medical records is provided: each step of reasoning is accompanied by a source identifier, including the knowledge graph path number and the medical record number, allowing clinicians to directly verify the reliability of the diagnostic basis and meet the high requirements for AI interpretability in high-risk medical decision-making scenarios.

[0034] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 The diagram below shows the overall architecture of the dual-source adaptive retrieval and generation system for atlases and medical records provided in this embodiment of the invention, illustrating the data flow and connection relationships between the various components of the system.

[0037] Figure 2 The system flowchart of the dual-source adaptive retrieval and generation method of atlas and medical records provided in the embodiments of the present invention shows the complete processing flow from the input of the patient's natural language chief complaint to the output of differential diagnosis suggestions, including the dictionary-driven medical entity linking stage, the dual-source parallel retrieval stage, the adaptive Z-score gating fusion stage, and the evidence tracing generation stage.

[0038] Figure 3 This is a schematic diagram of the local ontology structure of the dizziness-specific knowledge graph provided in an embodiment of the present invention. It shows an example of a multi-hop reasoning path topology with "positional vertigo" as the starting node, as well as illustrations of different types of entity nodes and relation edges. The medical codes involved in the figure are all based on the ICD-11 standard and are used for the standardized representation and reasoning support of diseases, symptoms, examinations and treatments. Detailed Implementation

[0039] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0040] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0042] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0043] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0044] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments, and the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0045] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0046] Example 1

[0047] In this embodiment 1, a dual-source adaptive retrieval and generation system for clinical decision-making in dizziness and vertigo is first provided, comprising: a dictionary-driven medical entity linker module, used to map the patient's colloquial complaints to a standardized set of clinical terminology entities; a dizziness-specific knowledge graph and multi-hop path reasoning engine module, used to perform multi-hop path reasoning with standardized clinical entities as anchor nodes, and output candidate reasoning paths with topological scores; an electronic medical record vector database and dense vector retrieval engine module, used to retrieve past medical records that are semantically similar to the patient's complaints in parallel, and output candidate medical records with semantic similarity scores; an adaptive Z-score gated fusion layer module, used to perform standardization and entropy-driven dynamic gating fusion on two sets of heterogeneous scores, and output the optimal subset of evidence; and an evidence tracing generation layer module, used to structurally assemble the optimal subset of evidence with patient metadata and system prompt templates, drive a large language model to perform chained reasoning, and output differential diagnosis suggestions with explicit evidence tracing.

[0048] In this embodiment, as Figures 1 to 3 As shown, the above system was used to implement a dual-source adaptive retrieval and generation method for clinical decision-making in dizziness and vertigo, including: mapping the patient's spoken complaints to a standardized set of clinical terminology entities; combining a dizziness and vertigo-specific knowledge graph, using standardized clinical entities as anchor nodes to perform multi-hop path reasoning, and outputting candidate reasoning paths with topological scores; parallelly retrieving previous medical records semantically similar to the patient's complaints, and outputting candidate medical records with semantic similarity scores; standardizing and fusion the topological scores and semantic similarity scores with entropy-driven dynamic gating, and outputting the optimal evidence subset; structuring and assembling the optimal evidence subset with patient metadata and system prompt templates, driving a large language model to perform chain reasoning, and outputting differential diagnosis suggestions with explicit evidence tracing.

[0049] A pre-constructed knowledge graph specifically for dizziness and vertigo is described. For a directed heterogeneous attribute graph, the node entity types must include at least the following five categories:

[0050] Positive symptom nodes: These represent vestibular-related symptoms that are clearly present in the patient, such as "rotational vertigo", "positional vertigo", "tinnitus", "hearing loss", "nausea and vomiting", etc.

[0051] Negative signs: These are signs that are clearly absent in the patient and play a key role in the differential diagnosis of exclusion, such as "no tinnitus", "no hearing loss", "no central oculomotor abnormalities", "no spontaneous nystagmus", etc.

[0052] Attack time pattern nodes: Characterize the duration of vertigo attacks, such as "duration < 1 minute", "lasts from several minutes to several hours", "lasts from several hours to 1 day", etc.

[0053] Triggering factors: These represent specific conditions that trigger or aggravate vertigo, such as "change in head position", "triggering by turning over", "triggering by waking up in the morning", "stress", etc.

[0054] Physical examination findings: positive or negative results of a specialist physical examination for vestibular function, such as "positive Dix-Hallpike test", "abnormal HINTS test", "positive Romberg test", etc.

[0055] The relation types of the directed edges include at least the following five categories:

[0056] "Confirmed diagnosis" relationship: This indicates that a set of clinical manifestations can be directly diagnosed as the target disease, and the clinical weight ranges from 0.9 to 1.0.

[0057] "Exclusion" relationship: The presence of a certain negative physical sign can exclude a specific disease, and the clinical weighting range is 0.9 to 1.0;

[0058] "Suggestive" relationship: Indicates that a certain clinical manifestation is highly suggestive of a specific disease, with a clinical weight ranging from 0.6 to 0.8;

[0059] "Accompanying" relationship: Characterizes common co-occurrence patterns among symptoms, with a clinical weighting range of 0.3 to 0.5;

[0060] "Potentially relevant" relationship: Characterizes a weak association between clinical manifestations and disease, with clinical weights ranging from 0.1 to 0.3.

[0061] Among them, the clinical weight value of the "exclusion" relationship is greater than that of the "accompanying" relationship, so as to reflect the priority of exclusion logic in differential diagnosis.

[0062] In this embodiment, the patient's spoken complaint text is obtained, and then mapped and standardized into a standardized set of clinical entities using an entity linking module based on a pre-set vertigo specialist dictionary. This set includes:

[0063] Receive the patient's natural language complaint text The natural language complaint text is the patient's original statement of dizziness and vertigo-related symptoms, described in a colloquial manner. Regarding the complaint text... Slide Window scanning, Take 1 to the preset maximum phrase length Obtain a collection of text fragments ,in A preset positive integer is used to control the maximum span of the window scan, to accommodate the variability in phrase lengths in Chinese medical terminology; a dense embedding model in the medical field is used for each text segment. Perform vectorization encoding to obtain fragment embedding vectors; obtain a pre-built specialty clinical dictionary, which contains... A set of standard vestibular medical terms, using a dense embedding model of the medical field for each dictionary term. Perform vectorization encoding; for each text segment With each dictionary term Calculate semantic alignment scores; extract all terms that meet the criteria. A standardized set of clinical entities is formed by using a preset high-confidence alignment threshold.

[0064] Based on the standardized clinical entity set, a dual-source knowledge retrieval mechanism is triggered in parallel, including:

[0065] On the one hand, topological evidence paths related to the standardized clinical entity set are retrieved in the pre-constructed dizziness-specific knowledge graph; on the other hand, the pre-constructed dizziness-specific knowledge graph... In China, with Each entity in the code is the starting anchor node. A multi-hop traversal is performed along the directed edges to obtain a set of candidate inference paths. For each reasoning path Calculate topology score ,in For relation type Preset clinical weights, For nodes Relevance to the set of anchor entities This serves as a path length normalization factor; on the other hand, it retrieves clinical medical record reference texts semantically similar to the standardized clinical entity set and the chief complaint text from the unstructured electronic medical record database; and in the pre-constructed electronic medical record vector database... China and Israel The mean pooling results of the embedding vectors of each entity in the original complaint The weighted concatenation result of the embedded vectors is used as the query vector, and approximate nearest neighbor retrieval is performed to obtain the query vector. Calculate the semantic similarity score for the most semantically similar medical records.

[0066] In this embodiment, the retrieved topological evidence paths and the clinical medical record reference text are relevance scored respectively, and the relevance scores are aligned using the Z-score normalization method to obtain the atlas confidence score and the medical record confidence score, including:

[0067] The knowledge graph path scoring set Calculate the mean and standard deviation And the score for each path is standardized using Z-score to obtain ;

[0068] The semantic similarity score set of the medical records Calculate the mean and standard deviation And perform Z-score standardization to obtain ;

[0069] Furthermore, through an adaptive gating mechanism, the fusion ratio weight between the map retrieval results and the medical record retrieval results is dynamically calculated based on the map confidence score and the medical record confidence score. A fusion prompt context is then constructed based on this fusion ratio weight, including:

[0070] Calculate the retrieved spectral sub-graph Shannon information entropy ,in For nodes Normalized importance in the subgraph, based on the information entropy and the entropy of the medical record retrieval results. Calculate dynamic gating parameters ;

[0071] Calculate the comprehensive integrated score ,according to Sort from highest to lowest, and select... These pieces of evidence constitute the optimal subset of evidence. .

[0072] Furthermore, the context of the fused cue words is input into a pre-trained large language model. Through generative reasoning of the large language model, auxiliary diagnostic results and exclusive differential evidence interpretations for dizziness and vertigo are output, including: the optimal subset of evidence. The graph reasoning path and similar medical records in the system are represented in text form; the diagnostic system prompt template is used. Standardized patient metadata Graph reasoning path text Similar medical record summaries Perform structured assembly to form ;Will Input a large language model, drive the model to perform chain-of-thought reasoning, analyze evidence segment by segment, compare with clinical manifestations, and make differential excluding; output differential diagnosis suggestions with explicit evidence tracing. and its reasoning and origin , Each step of reasoning is marked with the source of the evidence upon which it is based.

[0073] Example 2

[0074] This example describes a real outpatient scenario involving a 45-year-old female patient. The patient submits her natural language complaint to the vertigo clinic. Wei said, "For the past month, every time I get up in the morning or turn over in bed at night, I feel like the whole room is spinning wildly. It lasts for about 20 or 30 seconds and then stops. I don't have ringing in my ears or any difficulty hearing."

[0075] Phase 1: Dictionary-Driven Medical Entity Linking. The system receives the aforementioned chief complaint text through a dictionary-driven medical entity linker. Entity linker Slide Window scanning, Values ​​range from 1 to (This embodiment) ), obtain a collection of text fragments The medBERT model, a medical dense embedding model with an embedding dimension of 1024, was used to analyze each text segment. Perform vectorized encoding.

[0076] Entity linker acquires pre-built specialty clinical dictionaries The dictionary covers the field of vestibular medicine. Standardized terms, including but not limited to "rotational vertigo," "positional vertigo," "duration < 1 minute," "duration several minutes to several hours," "morning-induced," "turning over-induced," "no tinnitus," "no hearing loss," "tinnitus," "hearing decline," "nausea," "vomiting," and "headache," are vectorized and encoded using the same medical dense embedding model. For each... With each Calculate semantic alignment score .

[0077] Set a high confidence threshold Extract all terms that meet the criteria to form a standardized set of clinical entities:

[0078] .

[0079] At this point, "crazy spinning" has been precisely mapped to "rotational vertigo," "twenty to thirty seconds" has been precisely mapped to "duration < 1 minute," "no ringing in the ears" has been precisely mapped to the key negative sign "no tinnitus," and "unable to hear clearly" has been precisely mapped to the negative sign "no hearing loss." The physical linker will... The results are passed to the multi-hop path inference engine and the dense vector retrieval engine, respectively.

[0080] Phase Two: Dual-Source Parallel Search. The two search paths are initiated in parallel within the system.

[0081] Knowledge graph side: reasoning engine with Each entity serves as the starting anchor node in the pre-constructed knowledge graph specifically for dizziness and vertigo. Perform a multi-hop traversal along the directed edge in the middle, and determine the maximum number of hops. .

[0082] Retrieved candidate reasoning path set include:

[0083] path : (Posterior vertigo) → [Accompanying] → (Duration < 1 minute) → [Prompt] → (No tinnitus) → [Exclusion] → (No hearing loss) → [Prompt] → (BPPV - Posterior canal type), length 4 beats;

[0084] path : (Positive positional vertigo) → [Prompt] → (Triggered by turning over) → [Prompt] → (Triggered by waking up) → [Prompt] → (BPPV), with a length of 3 jumps;

[0085] path (Rotational vertigo) → [Accompanying] → (No hearing loss) → [Excluded] → (Meniere's disease), length 2 jumps. Calculate the topological score for each path.

[0086] by For example:

[0087] ;

[0088] by For example: ;

[0089] by For example: .

[0090] It should be understood that the specific values ​​mentioned above are only illustrative calculations for this embodiment, and the actual score depends on the calculation results of the specific graph structure and entity embedding model.

[0091] Electronic medical record side: Search engine with Calculate the joint query vector (where (For medical entity embedding functions), indexed by FAISS in the electronic medical record vector database Search The candidate medical records with the most similar semantics.

[0092] The search results are as follows: Medical records (EMR#1042, BPPV confirmed case, female, 43 years old), cosine similarity score 0.85; medical records (EMR#2057, BPPV confirmed case, female, 48 years old), Yu Xian's medical record (EMR#3891, confirmed BPPV case, male, 51 years old), cosine similarity score 0.79; medical records (EMR#1128, confirmed case of vestibular neuritis, female, 44 years old), cosine similarity score 0.71; medical records (EMR#5012, BPPV confirmed case, female, 47 years old), cosine similarity score 0.68.

[0093] Phase 3: Adaptive Z-score gating fusion.

[0094] Graph Path Scoring Set ,calculate , ,

[0095] ,

[0096] ,

[0097] .

[0098] Medical Record Semantic Similarity Scoring Set ,calculate , ,

[0099] ,

[0100] ,

[0101] ,

[0102] ,

[0103] .

[0104] The graph subgraph contains the set of nodes involved in the three reasoning paths. The graph information entropy is calculated after PageRank iteration. .

[0105] After the medical record retrieval results were normalized by softmax Medical record information entropy Dynamic gating parameters:

[0106] .

[0107] Should The medical interpretation of the value is as follows: because the path structure of the graph subgraph is compact and multiple paths converge at the BPPV node (low information entropy), while the medical record retrieval results score is high but the distribution is more dispersed than that of the graph side, the graph side is assigned a fusion weight of 0.59.

[0108] according to Calculate the comprehensive fusion score of each piece of evidence, and unify the atlas paths and similar medical records according to... Select after sorting = 5 pieces of evidence constitute the optimal subset of evidence .

[0109] Phase Four: Evidence Source Tracing and Generation. The evidence source tracing and generation layer will... The graph reasoning path and similar medical records in the data are textualized separately.

[0110] Diagnostic system prompt template (Including identification and exclusion logic guidelines), standardized patient metadata ("Patient, female, 45 years old, standard entity: rotational vertigo, positional vertigo, duration <1 minute, triggered in the morning, triggered by turning over, no tinnitus, no hearing loss") Graph reasoning path text ( and Natural language textual descriptions and similar medical record summaries (Medical Record) to The desensitized abstract is structured and assembled to form... Input the GPT-4o model to drive chain inference. The model output includes differential diagnostic suggestions with explicit evidence for tracing the source. and its reasoning and origin .

[0111] The output of this example is: Differential diagnosis: Benign paroxysmal positional vertigo (BPPV, posterior canal type), highly probable. The differential diagnosis criteria include:

[0112] Positional vertigo lasting less than 1 minute, triggered by turning over, and triggered in the morning are typical clinical features of BPPV (evidence source: KG pathway #P1, KG pathway #P2).

[0113] No tinnitus or hearing loss, effectively ruling out Meniere's disease (evidence source: KG pathway #P1: no tinnitus → [exclude] → no hearing loss → [exclude] → Meniere's disease);

[0114] Three previously diagnosed cases of BPPV provide empirical support (evidence sources: EMR#1042, EMR#2057, EMR#3891). Recommendation: Perform the Dix-Hallpike positional provocation test to further confirm the affected side and specific regulatory type, and to rule out other vestibular system disorders.

[0115] In terms of the technical effects of this embodiment, the system successfully achieved the following four key technical indicators:

[0116] Bridging the semantic gap: accurately mapping colloquial expressions such as "spinning wildly," "twenty or thirty seconds," and "no ringing in the ears" to standardized terms such as "rotational vertigo," "duration < 1 minute," and "no tinnitus."

[0117] Parallel acquisition of dual-source evidence: Three high-quality reasoning paths containing exclusionary logic are retrieved from the knowledge graph, while five candidate medical records sorted by semantic similarity are retrieved from the medical record vector database.

[0118] Heterogeneous evidence scale unification and fusion: After Z-score standardization, the spectral path At 1.15 and medical records With 1.16 At a comparable scale, and through entropy-driven gating A fair weighted fusion was performed;

[0119] Evidence tracing: Each inference in the output is marked with a corresponding KG path number or EMR number, achieving complete traceability of the diagnostic conclusion.

[0120] Example 3

[0121] This embodiment describes different parameter configuration schemes for the same system to demonstrate the adaptability of the method of the present invention in different application scenarios. In this embodiment, the system is deployed for the initial screening scenario of vestibular diseases in primary community health service centers. Considering that the electronic medical record database in primary scenarios is relatively small (usually only a few thousand records) and computing resources are limited (usually only equipped with CPU servers), the system parameters are adjusted for adaptability as follows:

[0122] slide Maximum window length Set to 3 (a reduction from 5 in Example 1 to reduce the computational cost of window enumeration);

[0123] Entity link semantic alignment threshold The value was set to 0.80 (a reduction from 0.85 in Example 1 to improve recall and ensure that key entities are not missed when there are more variations in non-standardized spoken language at the grassroots level).

[0124] Dual-source retrieval The quantity is set to 10 (an increase from 5 in Example 1, to compensate for the potential insufficiency of information in a single medical record in a small-scale medical record database).

[0125] Pooling weighting coefficients on the medical record side The value was set to 0.7 (an increase from 0.6 in Example 1 to enhance the guiding role of standardized clinical entity anchors and reduce the impact of potential noise in the original spoken text).

[0126] Maximum number of hops in knowledge graph path reasoning The value is set to 3 (a reduction from 4 in Example 1, to meet the needs of simplified diagnostic pathways and rapid response in primary screening scenarios at the grassroots level).

[0127] The dense embedding model employs a lightweight variant with an embedding dimension of 768 to maintain acceptable retrieval latency in CPU environments. After the above parameter adaptation, the system can complete the entire process from chief complaint input to initial screening suggestion output within 3 seconds in a primary care deployment environment (approximately 1.5 seconds in the GPU deployment environment of the tertiary hospital in Example 1), meeting the high-throughput and rapid triage needs of primary care clinics.

[0128] It should be understood that the above parameter configuration scheme is only another exemplary implementation of the present invention. In actual deployment, each parameter can be independently adjusted according to the data scale, computing resources and clinical needs of the specific scenario.

[0129] It should be noted that the parameter settings in the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the claims. The following provides recommended value ranges for each key parameter in general scenarios: sliding... Maximum window length The recommended value range is 3 to 6. Its selection depends on the phrase length distribution characteristics of the target language. Key terms in Chinese vestibular medical terms are mainly 2 to 6 characters long. Therefore, this range can effectively cover the window capture needs of most clinically relevant terms.

[0130] Entity link semantic alignment confidence threshold The recommended value range is 0.80 to 0.95. This value reflects the classic trade-off between precision and recall. For the precise diagnosis scenario in tertiary hospitals, it is recommended to use 0.88 to 0.95 to prioritize link accuracy. For the primary screening scenario in primary care, it is recommended to use 0.80 to 0.87 to ensure broader coverage of spoken language.

[0131] Dual-source retrieval The recommended value for the number is 5 to 20, depending on the effective context window length of the downstream large language model and the need for diversity of input evidence;

[0132] The recommended range for the dimension of medical dense embedding vectors is 768 to 4096, depending on the specific specifications of the selected embedding model and the computational resource constraints of the deployment environment.

[0133] Clinical weights of each relationship type In the above categories, the recommended value range is 0.9 to 1.0 for "exclusion", 0.8 to 1.0 for "diagnosis basis", 0.6 to 0.8 for "hints", 0.3 to 0.5 for "accompanying", and 0.1 to 0.3 for "possibly related".

[0134] Furthermore, the explanations of technical terms in the above embodiments include, but are not limited to, the specific model names used (such as GPT-4o, FAISS, HNSW, etc.), embedding dimension values, and retrieval. The quantities, threshold parameter values, weight coefficient values, maximum number of hops, etc., are merely illustrative examples and should not be construed as limiting the scope of protection of the claims of this invention.

[0135] Example 4

[0136] This embodiment 4 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo, as described above. The method includes:

[0137] Mapping patients' verbal complaints to a standardized set of clinical terminology entities;

[0138] By combining a knowledge graph specifically for dizziness and vertigo, multi-hop path reasoning is performed with standardized clinical entities as anchor nodes, and candidate reasoning paths with topological scores are output.

[0139] Parallel retrieval of previous medical records that are semantically similar to the patient's chief complaint, and output of candidate medical records with semantic similarity scores;

[0140] The topological score and semantic similarity score are standardized and fused with entropy-driven dynamic gating to output the optimal subset of evidence;

[0141] The optimal subset of evidence is structurally assembled with patient metadata and system prompt templates to drive a large language model to perform chained reasoning and output differential diagnosis suggestions with explicit evidence tracing.

[0142] Example 5

[0143] This embodiment 5 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the dual-source adaptive retrieval and generation method of atlas and medical records for clinical decision-making in dizziness and vertigo as described above. The method includes:

[0144] Mapping patients' verbal complaints to a standardized set of clinical terminology entities;

[0145] By combining a knowledge graph specifically for dizziness and vertigo, multi-hop path reasoning is performed with standardized clinical entities as anchor nodes, and candidate reasoning paths with topological scores are output.

[0146] Parallel retrieval of previous medical records that are semantically similar to the patient's chief complaint, and output of candidate medical records with semantic similarity scores;

[0147] The topological score and semantic similarity score are standardized and fused with entropy-driven dynamic gating to output the optimal subset of evidence;

[0148] The optimal subset of evidence is structurally assembled with patient metadata and system prompt templates to drive a large language model to perform chained reasoning and output differential diagnosis suggestions with explicit evidence tracing.

[0149] Example 6

[0150] This embodiment 6 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the dual-source adaptive retrieval and generation method of atlas and medical records for clinical decision-making in dizziness and vertigo as described above. The method includes:

[0151] Mapping patients' verbal complaints to a standardized set of clinical terminology entities;

[0152] By combining a knowledge graph specifically for dizziness and vertigo, multi-hop path reasoning is performed with standardized clinical entities as anchor nodes, and candidate reasoning paths with topological scores are output.

[0153] Parallel retrieval of previous medical records that are semantically similar to the patient's chief complaint, and output of candidate medical records with semantic similarity scores;

[0154] The topological score and semantic similarity score are standardized and fused with entropy-driven dynamic gating to output the optimal subset of evidence;

[0155] The optimal subset of evidence is structurally assembled with patient metadata and system prompt templates to drive a large language model to perform chained reasoning and output differential diagnosis suggestions with explicit evidence tracing.

[0156] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0157] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0158] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0159] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0160] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A dual-source adaptive retrieval and generation method for atlas and medical records in clinical decision-making for dizziness and vertigo, characterized in that, include: Mapping patients' verbal complaints to a standardized set of clinical entities; Based on the standardized clinical entity set, multi-hop path reasoning is performed in the dizziness and vertigo-specific knowledge graph with each entity as an anchor node to obtain candidate reasoning paths and calculate the corresponding topology scores. Parallel retrieval of past medical records semantically related to the patient's chief complaint and the standardized clinical entity set is performed to obtain candidate medical records and calculate the corresponding semantic similarity score; The topological score and the semantic similarity score are both standardized. Dynamic fusion weights are determined based on the information entropy of the evidence set on the graph side and the information entropy of the evidence set on the medical record side, and a comprehensive fusion score is calculated accordingly; the optimal subset of evidence is selected based on the comprehensive fusion score. The optimal subset of evidence is structurally assembled with patient metadata and system prompt templates, driving the large language model to perform chain reasoning and output differential diagnosis suggestions with explicit evidence tracing.

2. The dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo according to claim 1, characterized in that, A pre-constructed knowledge graph specifically for dizziness and vertigo is provided. This knowledge graph is a directed heterogeneous attribute graph, and its node entity types include at least the following five categories: Positive symptom nodes, negative physical sign nodes, onset time pattern nodes, precipitating factor nodes, and physical examination findings nodes; the directed edge relationship types include at least the following five categories: confirmatory relationship, exclusion relationship, suggestive relationship, accompanying relationship, and possibly related relationship; among them, the clinical weight value of the exclusion relationship is greater than the clinical weight value of the accompanying relationship, so as to reflect the priority of exclusion logic in differential diagnosis.

3. The dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo according to claim 1, characterized in that, The patient's spoken complaint text is obtained, and then mapped and standardized into a standardized set of clinical entities using an entity linking module based on a pre-built vertigo specialist dictionary. This set includes: The system receives the patient's natural language complaint text, which is the patient's original description of dizziness and vertigo-related symptoms in a colloquial manner; it performs a sliding window scan on the complaint text to obtain a set of text fragments, controlling the maximum span of the window scan to accommodate the variability in phrase length in Chinese medical terminology; it uses a dense embedding model in the medical domain to vectorize each text fragment, obtaining a fragment embedding vector; and it acquires a pre-built specialty clinical dictionary, which includes... A set of standard vestibular medical terms is defined, and each dictionary term is vectorized using the dense embedding model of the medical field. For each text segment and each dictionary term, a semantic alignment score is calculated. All terms that meet the criteria are extracted to form a standardized set of clinical entities.

4. The dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo according to claim 1, characterized in that, Based on the standardized clinical entity set, a parallel dual-source knowledge retrieval mechanism is triggered, including: retrieving topological evidence paths related to the standardized clinical entity set in a pre-constructed dizziness-specific knowledge graph; performing multi-hop traversal along directed edges in the pre-constructed dizziness-specific knowledge graph, with each entity as the starting anchor node, to obtain a set of candidate reasoning paths, and calculating a topological score for each reasoning path; retrieving clinical medical record reference texts semantically similar to the standardized clinical entity set and the chief complaint text in an unstructured electronic medical record database; and using the weighted concatenation result of the mean pooling result of the embedding vectors of each entity and the embedding vector of the original chief complaint as the query vector in a pre-constructed electronic medical record vector database, obtaining the semantically most similar medical record through approximate nearest neighbor retrieval, and calculating a semantic similarity score.

5. The dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo according to claim 1, characterized in that, The context of the fused prompt words is input into a pre-trained large language model. Through generative reasoning of the large language model, auxiliary diagnostic results and exclusive differential evidence interpretations for dizziness and vertigo are output, including: textualizing the graph reasoning paths and similar medical records in the optimal evidence subset; structurally assembling the diagnostic system prompt template, standardized patient metadata, graph reasoning path text, and similar medical record summaries and inputting them into the large language model to drive the model to perform reasoning, analyze the evidence segment by segment, compare with clinical manifestations, and perform differential exclusion; and output differential diagnostic suggestions with explicit evidence tracing and their reasoning tracing.

6. The dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo according to claim 1, characterized in that, Through an adaptive gating mechanism, based on the confidence scores of the atlas and medical records, the fusion ratio weight of the atlas retrieval results and the medical record retrieval results is dynamically calculated. A fusion prompt context is then constructed based on this fusion ratio weight, including: calculating the Shannon information entropy of the retrieved atlas subgraphs; calculating dynamic gating parameters based on the information entropy and the entropy of the medical record retrieval results; calculating a comprehensive fusion score; and selecting results from high to low. These pieces of evidence constitute the optimal subset of evidence.

7. A dual-source adaptive retrieval and generation system for atlas and medical records for clinical decision-making in dizziness and vertigo, characterized in that, include: The dictionary-driven medical entity linker module is used to map patients' spoken complaints to a standardized set of clinical terminology entities. The Dizziness and Vertigo Specialized Knowledge Graph and Multi-hop Path Reasoning Engine Module is used to perform multi-hop path reasoning with standardized clinical entities as anchor nodes and output candidate reasoning paths with topological scores. The electronic medical record vector database and dense vector retrieval engine module are used to retrieve past medical records that are semantically similar to the patient's chief complaint in parallel, and output candidate medical records with semantic similarity scores. The adaptive Z-score gated fusion layer module is used to standardize and entropy-driven dynamic gating fusion of two sets of heterogeneous scores, and output the optimal subset of evidence. The evidence tracing generation layer module is used to structurally assemble the optimal subset of evidence with patient metadata and system prompt templates, drive the large language model to perform chain reasoning, and output differential diagnosis suggestions with explicit evidence tracing.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the dual-source adaptive retrieval and generation method for atlas and medical records for clinical decision-making in dizziness and vertigo as described in any one of claims 1-6.

9. A computer device, characterized in that, The method includes a memory and a processor, the processor and the memory communicating with each other, the memory storing program instructions executable by the processor, and the processor calling the program instructions to execute the dual-source adaptive retrieval and generation method for clinical decision-making in dizziness and vertigo as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the dual-source adaptive retrieval and generation method of atlas and medical records for clinical decision-making in dizziness and vertigo as described in any one of claims 1-6.

Images